Selected Work

Projects

Seven systems built at the intersection of machine learning and engineering. Each one designed for production environments, not sandboxes.

NexusOS

Python FastAPI LangGraph Docker Redis

AI-native operating system interface with modular agent architecture and real-time orchestration.

Unified execution surface for concurrent AI agents — stateless compute with Redis stream-backed persistent state and hot-swap model routing.

Engineering Challenge

Lock-free coordination across agents with shared memory required a custom linearizability protocol over Redis pub/sub.

OmniAgent

Python LangChain Anthropic Celery PostgreSQL

Multi-agent autonomous task execution framework with cross-agent communication and persistent context.

Planner agent decomposes goals into subtask graphs; executor agents handle individual nodes via Celery-distributed workers with full tool access.

Engineering Challenge

Coherent shared context across parallel agents required a versioned memory store with optimistic concurrency and vector clock conflict resolution.

OmniRAG

Python Qdrant LlamaIndex FastAPI OpenAI

Production-grade retrieval-augmented generation with hybrid sparse-dense search and multi-document reasoning.

Hybrid BM25 + dense vector retrieval over heterogeneous corpora, re-ranked by a cross-encoder before generation with structured citations.

Engineering Challenge

Multi-document reasoning over long contexts solved with hierarchical summarisation and sentence-level evidence extraction.

InfraOS AI

Go Terraform Kubernetes OpenAI Ansible

Infrastructure automation platform driven by natural language — declarative cloud operations via LLM-generated IaC.

Natural language infrastructure intent translated to validated Terraform and Ansible, gated by policy-as-code dry-runs across AWS, GCP, and Azure.

Engineering Challenge

Deterministic infrastructure generation from ambiguous language required constrained generation with formal schema validation and per-step rollback capability.

CodeForge AI

Python TreeSitter Claude API FastAPI TypeScript

AI-powered code generation and review system with multi-language AST analysis and context-aware completions.

TreeSitter AST analysis across 12 languages feeds structured code representations to Claude for generation, inline diff review, and language-specific linting.

Engineering Challenge

Context window compression used a graph-based dependency resolver that prunes semantically distant nodes while preserving critical call chains.

CyberShield AI

Python PyTorch Elasticsearch Kafka FastAPI

Autonomous threat detection and response system using neural anomaly detection at network scale.

Kafka-ingested network telemetry scored by an LSTM autoencoder; high-confidence reconstruction anomalies trigger automated response playbooks with tunable threshold calibration.

Engineering Challenge

Sub-200ms detection at 100k events/second achieved with batched GPU inference, INT8 quantisation, and a ring buffer for sliding-window feature extraction.

InsightFlow AI

Python Airflow dbt BigQuery LangChain

Intelligent data pipeline with AI-driven analytics — automated insight generation from multi-source business data.

Airflow orchestrates multi-source ingestion, dbt transforms to BigQuery, and a LangChain analyst generates scheduled narrative reports, anomaly summaries, and trend analysis.

Engineering Challenge

Grounded insight generation required a tool-augmented approach where the LLM executes SQL against BigQuery rather than reasoning over raw data.